ARTIFACTS REMOVING FROM EEG SIGNALS BY

ICA ALGORITHMS
A.SRINIVASULU 1
Department of ECE,
MITS, Madanapalli,
Email id: akulamtech.815@gmail.com
Asso.prof.M.SRINATH REDDY2
Department of ECE,
MITS, Madanapalli,

ABSTRACT: Recent advances in computer hardware and signal processing have now a way to
communicate with the outside world, but even with the last modern techniques, such systems still suffer
communication Recent advances in computer hardware and signal processing have made possible the
use of EEG signals or “brain waves” fo forr communication between humans and computers. Locked-in Locked
patient’s rates on the order of 2-3 3 tasks/minute. In addition, existing systems are not likely to be designed
with flexibility in mind, leading to slow systems that are difficult to improve.
The
he EEG is composed of electrical potentials arising from several sources. Each source (including
separate neural clusters, blink artifact, or pulse artifact) projects a unique topography onto the scalp
scalp-'scalp
maps'. These maps are mixed according to the pr principle
inciple of linear superposition. Here we attempt a
Independent component analysis (ICA) of different algorithms to reverse the superposition by separating
the EEG into mutually independent scalp maps, later removing their noise by set the threshold level a and
finally we classify the five mental tasks through the use of the electroencephalogram (EEG) by the neural
network technique.
Keywords- EEG database, Different Independent component analysis (ICA) Algorithms,, MAT LAB,

1. INTRODUCTION[5]: 2) Signal Pre-Processing:

Processing: once the signals are acquired, it is
necessary to clean them.
Brain Computer Interface Technology: 3) Signal Classification: once the signals are cleaned, they will
Present BCI’s use EEG activity recorded at the scalp to be processed and classified to find out which kind of mental task
control cursor movement, select letters or icons, or operate a the subject is performing
neuroprosthesis. The central element in each BCI is a translation 4) Computer Interaction: once the signals are classified, they
.4)
algorithm that converts electrophysiological input from the user will be used by an appropriate algorithm for the development of
into output that controls external devices. BCI operation depends a certain application.
on effective interaction between two adaptive controllers: the
user who encodes his or her commands in the
electrophysiological input provided ed to the BCI, and the computer
which recognizes the command contained in the input and
expresses them in the device control.
Current BCI’s have maximum information transfer rates of
5-25 bits/min. Achievement of greater speed and accuracy
depends on improvements in: • Signal acquisition: methods for
increasing signal-to-noise ratio (SNR), signal--to interference
ratio (S/I)) as well as optimally combining spatial and temporal
Information. • Single trial analysis: overcoming noise and
interference in order to avoid averaging and maximize bit rate.
• Co-learning:
learning: jointly optimizing combined man man-machine
system and taking Advantage of feedback.
• Experimental paradigms for interpretable readable signals:
mapping the task to the brain state of the user (or vice versa).
• Understanding algorithms and models within the context of
the neurobiology: Building predictive models having
neurophysiologically meaningful parameters and incorporating
physically and biologically meaningful priors.
The common structure of a Brain Computer Inte Interface is
shown in fig. (1) And the following are:
1) Signal Acquisition: the EEG signals are obtained from the
brain through invasive or non-invasive
invasive methods (for example,
electrodes). After, the signal is amplified and sampled. Fig. (1) Structure of Brain Computer Interface
2. METHOD[5]: Algorithms for ICA
EEG Signal Pre – Processing: There are a lot of kinds of algorithm for ICA. Some of them
One of the main problems in the automated EEG analysis are follows:
is the detection of the different kinds of interference waveforms -The algorithm relies on batch computations minimizing or
(artifacts) added to the EEG signal during the recording sessions. maximizing contrast functions based on higher-order cumulate.
These interference waveforms, the artifacts, are any recorded - The algorithm based on stochastic gradient methods, which
electrical potentials not originated in brain. There are four main may have implementations in neural networks.
sources of artifacts
emission:
1. EEG equipment. A. Fast-ICA [1]
2. Electrical interference external to the subject and recording The Fast-ICA algorithm belongs to the family of fix-point
system. algorithms for ICA, which is based on the iteration to search for
3. The leads and the electrodes. the maximum of the non-Gaussianity of variables.
4. The subject her/himself: normal electrical activity from the
Fixed-point algorithm
heart, eye blinking, eyes movement, and muscles in general.
Optimizes negentropy (negative entropy) or kurtosis to
Removing EEG artifacts by ICA algorithms: measure non-gaussianity.
Independent component analysis (ICA) is a relatively
Negentropy of X is defined by
recent method for blind source separation (BSS), which has J(X) = H(Xgauss) − H(X)
shown to outperform the classical principal component analysis
Good statistical properties, but computationally
(PCA) in many applications. In particular, it has been applied for difficult.
the extraction of ocular artifacts from the EEG, where principal
Approximations of negentropy
PCA could not separate eye artifacts from brain signals, JG(x) = (E{G(x)}−E{G(xgauss)})2 where G(.) is a non
especially when they have comparable amplitudes. quadratic function
This method presents some advantages compared to other 1 u2
rejection methods, such as: G1 (u ) = log cosh a1u G2 (u ) = − exp( − )
1. ICA separates EEG signals including artifacts into a1 2
independent components based on the characteristics of the data,
without relying on the availability of one or more “clean” 1 4
reference channels for each type of artifact. This avoids the
G3 (u ) = u
4
problem of mutual contamination between regressing and
The steps of Fast ICA algorithm
regressed channels. 1. Centering the input observation signal x
2. ICA-based artifact removal can preserve all of the recorded 2. Whitening the centered signal x
trials, a crucial advantage over rejection-based methods when
3. Initializing weight matrix w, and set convergence error ε
limited data are available, or when blinks and muscle movements
4. Update weight matrix w. Using following iterative formula
occur too frequently, as in some subject groups.
3. Unlike regression methods, ICA-based artifact removal can
preserve data at all scalp channels, including frontal and Wi + = Wi − µ [ E{ xg (Wi T x )} − β Wi ] /[ E{ g ' (Wi T x )} − β ]
particular sites.
Nevertheless, it is important to keep in mind that it also has T T
some inherent where β = E{W i xg (W i x)} ,
Limitations, such as:
1. ICA can decompose at most N sources from N scalp
electrodes. Usually, the effective number of temporally- 5. Normalize weight vector w: + + T
W i
= W i
/ W i
independent signals contributing to the scalp EEG is unknown,
and it is likely that observed brain activity arises from more 6. If | Wk +1 − Wk |> ε algorithm is not reach convergence,
Physically separable effective sources than the available number repeat steps (4) and (5);
of EEG Electrodes. 7. Get the separation matrix w
2. The assumption of temporal independence used by ICA
cannot be satisfied when the training data set is too small, or
when separate topographically distinguishable Phenomena
always occur concurrently in the data. In the latter case,
simulations show that ICA may derive a component accounting B. JADE [2]
for their joint occurrence, plus Separate components accounting Joint Approximate Diagonalization of Eigenmatrices
for their periods of solo activation. Such confounds imply that à Built on cumulants - based contrast function.
converging behavioral or other evidence must be obtained before
concluding that spatio-temporally overlapping ICA components
Fourth order cross- cumulants for zero-mean random
measure variables xi, xj , xk, xl, the cross-cumulant is defined as
Neuron-physiologically or functionally distinct activities. Cum(xi, xj , xk, xl) = E[xixjxkxl] - E[xixj ]E[xkxl]-
3. ICA assumes that the physical sources of artifactual and E[xixk]E[xjxl] -E[xixl]E[xjxk]
neural activity contributing to EEG signals are spatially
Auto cumulant-
stationary through time. In general, there is no reason to believe Cum(xixixixi) = E[xixixixi]-
that cerebral and artifactual sources in the spontaneous EEG E[xixi]E[xixi]-E[xixi]E[xixi]-E[xixi]E[xixi]
necessarily remain stationary over time or occurrences. =E [xi4] – 3 (E[xi2])2
4. The fact that this method needs more computations compared kurt(X)= E(X4) − 3(E(X2))2
to a rejection approach, together with the inherently real-time
kurt(X) = 0 if X Gaussian, < 0 if sub and > 0 if super
nature of the EEG Brain Computer Interface makes its use a
jointly diagonalizing eigenmatrices of the kurtosis
more difficult alternative. .Steps in JADE

Initialization. Estimate a whitening matrix Wˆ and set
Z = Wˆ X.

Form statistics. Estimate a maximal set {Qˆ iZ } EEG Data Set: Mental Tasks
of cumulant matrices. Ø Resting task (Baseline)

Optimize an orthogonal contrast. Find the rotation Ø Imagined letter composing
Ø Mental multiplication
matrix Vˆ such that the cumulant matrices are as
Ø Visualized counting
diagonal as possible, that is, solve Ø Geometric object rotation
Vˆ = arg min Σi off (V T Qˆ iZV )

Separate. Estimate A as Aˆ = VˆWˆ −1
and/or estimate the components as
Sˆ = Aˆ −1 X = Vˆ T Z
C. INFOMAX ALGORITHM [3]

Principle

Sources are assumed independent.
⇒ they don’t have mutual information Fig.(2) Electrode placement on to the scalp

Therefore, minimizing the mutual information in
observed signals will lead to the separated signals. Electrodes : C3, C4, P3, P4, O1, O2 and EOG
Mutual Information >=0 and zero if and only if the
variables are statistically independent

3. RESULTS:
MEASURED SIGNALS
RAW EEG
600

500

400

Information Measures 300

Information is given by,
X (t )

Entropy: 1 200
I ( X ) = ∫ p( x) log dx
p ( x) 1 100
Joint Entropy: I ( X , Y ) = ∫ p( x, y ) log dxdy
p ( x, y ) 0
Mutual Information,
M(X,Y)= I(X) + I(Y) – I(X,Y) -100
0 500 1000 1500 2000 2500
NUMBER OF SAMPLES
Fig. (3) Raw EEG signals

FAST ICA
Nonlinearity: log(cosh(y))
No. of iterations: 100
Convergence error: 10e-300
10e
Minimizing Mutual Information ≈ Maximizing Joint Entropy/ INDEPENDENT COMPONENT USING FASTICA
Likelihood /Network Entropy 70

60
D. EXTENDED INFOMAX [2]
50

Extension of Infomax 40

This preserves the ICA architecture of Infomax
S (t)

algorithm 30
T T
∆W = [1 − K tanh(u )u − uu ]W 20

K = 1 : Supergaussian 10
K = -11 : Subgaussian
0

-10
0 500 1000 1500 2000 2500
NUMBER OF SAMPLES
Fig.(4) Pure EEG signal by Fast ICA algorithm
JADE

INDEPENDENT COMPONENTS USING JADE

70
RECONSTRUCTED EEG
60 300

50
250
40

200
S(t)

30

20
150
10

X(t)
0 100

-10
0 500 1000 1500 2000 2500 50
NUMBER OF SAMPLES
Fig.(5) Pure EEG signal by JADE ICA algorithm
0

INFOMAX
Convergence eeror =1e-3 -50
Transformation function =logistic sigmoid = 1 0 500 1000 1500 2000 2500
NUMBER OF SAMPLES
Number of iterations: 1 + e −u
512
Fig(8) pure EEG signals without artifacts
INDEPENDENT COMPONENTS USING INFOMAX
140
4. CONCLUSION
120 ICA is the central topic in this paper. EEG signals will
maintain the similarity in their patterns when subject is
100 performing the mental task. BCI systems using EEG as
control signal suffers from the artifact problem. The
80 traditional methods applied for remove artifacts can only
compromise between eliminating artifacts and protecting
S (t )

60 useful signals so that the result is not very satisfying.

However, ICA method can protect the useful signals as well
40
as obviously weaken even entirely remove the artifacts in
multi channel EEG signals, this characteristic of ICA is the
20
key to get stable EEG patterns which can be used for mental
task classification.
0

-20
0 500 1000 1500 2000 2500 5. REFERENCE:
NUMBER OF SAMPLES 1. Aapo Hyvarinen., “fast and Robust fixed point algorithm
Fig.(6) Pure EEG signal by Infomax ICA algorithm for independent component analysis,” IEEE transactions
on Neural Networks, Vol.10, pp.626-634, 1999.
EXTENDED INFOMAX 2. Amari , and Cichocki., “A new learning algorithm for
blind source separation,” Advances in Neural Information
Convergence error: 1e-3 Processing, MIT press, pp.757-763, vol.8, 1996.
Number of iterations=512 3. A.J.Bell and T.J.Sejnowski, “An information-maximization
Nonlinearity = tanh(u) approach to blind separation and blind deconvolution,”
Neural Computation, vol.7, pp.1129-1159, 1995.
INDEPENDENT COMPONENT USING EXTENDED INFOMAX
140 4. T.W.Lee and T.J.Sejnowski, “Independent component
analysis for sub-gaussian and super-gaussian mixtures,”
120 Proceedings of 4th joint symposiums on Neural
computations, vol.7, pp.132-139, 1997.
100
5. Sarah M. Hosni, and Mahmoud E.gadallah.,
80 “Classification of EEG signals using different feature
extraction techniques for mental-task BCI,” Ain Shams
university, Cairo, Egypt.
S (t)

60
6. Hyv arinen A., Karhunen J., Oja E.: Independent
40 Component Analysis. Wiley {
Interscience, 2001.
20
7. M__ka S.: Numerick_e metody algebry. SNTL, 1985.
0 8. Nev_s__malov_a S., _ Sonka K. a spol.: Poruchy sp_anku
a bd_en__. Maxdorf s.r.o., 1997
-20 9. Data downloaded from:
0 500 1000 1500 2000 2500 http://www.cs.colostate.edu/~anderson.
NUMBER OF SAMPLES
Fig. (7) Pure EEG signal by Extended Infomax ICA algorithm
RECONSTRUCTED EEG